WO2024026990A1 - Automatic iterative training method, system and device for recognition model, and storage medium - Google Patents

Abstract

Provided in the present disclosure are an automatic iterative training method, system and device for a recognition model, and a computer-readable storage medium. By means of the technical solution provided in the present disclosure, training data required for iteration of a recognition model can be automatically generated by means of recognition results and verification results that are obtained between different rounds of training, and the training data is used to automatically update and iterate the recognition model, so as to avoid the situations of a recognition model aging and failing to meet a recognition requirement, such that the manpower cost on the operation, maintenance and upgrade of a recognition model is saved on, thus having a value of popularization.

Description

Automatic iterative training methods, systems, devices and storage media for recognition models Technical field

The present disclosure relates to the technical field of model training. Specifically, an automatic iterative training method, system, equipment and storage medium for recognition models are disclosed.

Background technique

In the field of model recognition technology, the trained recognition model is usually used to identify the target object to be recognized to obtain the recognition result. With the continuous improvement of the demand for model recognition accuracy and the continuous updating and iteration of the types of recognition objects, it is often necessary to retrain the obtained recognition models to meet the increasing changes in recognition requirements.

Contents of the invention

The present disclosure provides an automatic iterative training method, system, device and storage medium for a recognition model. Specifically, the first aspect of the present disclosure provides an automatic iterative training method for a recognition model, including the following steps:

Perform training operations on the recognition model based on the training data required for the current training round;

Perform an inference operation on the recognition model that has completed the current training round to obtain the first recognition result containing pre-annotation information;

Verify the first identification result to obtain the second identification result including verification information;

Comparing the first recognition result of the previous training round of the recognition model with the second recognition result of the current training round to generate training data required for the next training round;

Repeat the above steps to achieve automatic iterative training of the recognition model.

In a possible implementation of the first aspect above, a preset model iteration time is separated between two adjacent training rounds.

In a possible implementation of the above first aspect, the recognition model is used to recognize an image to be recognized that contains a target object to obtain identification information of the target object;

The training data contains images to be recognized.

In a possible implementation of the first aspect above, performing a training operation on the recognition model includes the following steps:

Divide the image to be recognized to obtain the corresponding foreground area;

Perform first preset processing on the foreground area to obtain the minimum circumscribed rectangular marked area containing the target object;

Perform second preset processing on the image to be recognized to obtain a preferred training set containing the minimum circumscribed rectangle marked area;

Train the recognition model through the optimized training set to generate a pre-trained model;

The pre-trained model is optimized so that the pre-trained model that meets the preset evaluation conditions is used as the recognition model to complete the current training round.

In a possible implementation of the above first aspect, performing the first preset processing on the foreground area includes the following steps:

Obtain the edge line of the foreground object in the foreground area to obtain the angle of the foreground object in the image to be recognized;

Rotate the foreground area according to the angle so that the foreground object is in a vertical or horizontal direction relative to the image to be recognized;

Obtain the enclosing rectangular marking frame containing the target in the rotated foreground area as the minimum enclosing rectangular marking area.

In a possible implementation of the first aspect above, performing the second preset processing on the image to be recognized includes the following steps:

Perform sliding window cropping on the image to be identified according to the preset window size to obtain several cropped images that overlap with the minimum circumscribed rectangle marked area;

Perform data augmentation on cropped images to generate optimal training sets.

In a possible implementation of the first aspect above, optimizing the pre-trained model includes the following steps:

The recognition results of the pre-trained model are post-processed using non-maximum value merging.

A second aspect of the present disclosure provides an automatic iterative training system for a recognition model, which is applied to the automatic iterative training method for a recognition model provided by the first aspect;

An automated iterative training system for recognition models includes:

The training unit is used to perform training operations on the recognition model based on the training data required for the current training round;

An inference unit, used to perform inference operations on the recognition model that has completed the current training round to obtain the first recognition result containing pre-annotation information;

A verification unit, used to verify the first identification result to obtain the second identification result containing verification information;

A generation unit configured to compare the first recognition result of the previous training round of the recognition model with the second recognition result of the current training round to generate training data required for the next training round;

Iterative unit is used to repeat the above steps to achieve automatic iterative training of the recognition model.

A third aspect of the present disclosure provides an automatic iterative training device for a recognition model, including:

Memory, used to store computer programs;

The processor is configured to implement the automatic iterative training method of the recognition model provided in the first aspect when executing the computer program.

A fourth aspect of the present disclosure provides a computer-readable storage medium, which is characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, it implements automatic iteration of the recognition model provided in the first aspect. Training methods.

Compared with the prior art, the present disclosure has the following beneficial effects:

Through the technical solution proposed in this disclosure, the training data required for the iteration of the recognition model can be automatically generated through the recognition results and verification results obtained between different training rounds, and these training data can be used to automatically update and iterate the recognition model to avoid The recognition model is aging and cannot meet the recognition needs. At the same time, it saves the labor cost of operation and maintenance upgrade of the recognition model, which has scalable value.

Description of the drawings

Other features, objects and advantages of the present invention will become more apparent by reading the detailed description of the non-limiting embodiments with reference to the following drawings:

Figure 1 shows a schematic flow chart of an automatic iterative training method for a recognition model according to an embodiment of the present disclosure;

Figure 2 shows a schematic flowchart of performing a training operation on a recognition model according to an embodiment of the present disclosure;

Figure 3 shows a schematic flowchart of performing first preset processing on the foreground area according to an embodiment of the present disclosure;

Figure 4 shows a schematic flow chart of a second preset processing of an image to be recognized according to an embodiment of the present disclosure;

Figure 5 shows a schematic structural diagram of an automatic iterative training system for recognition models according to an embodiment of the present disclosure;

Figure 6 shows a schematic structural diagram of an electronic device according to an embodiment of the present disclosure;

FIG. 7 shows a schematic structural diagram of a computer-readable storage medium according to an embodiment of the present disclosure.

Specific implementation methods

The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

As used herein, the term "include" and its variations mean an open inclusion, ie, "including but not limited to." Unless otherwise stated, the term "or" means "and/or". The term "based on" means "at least regionally based on". The terms "one example embodiment" and "an embodiment" mean "at least one example embodiment." The term "another embodiment" means "at least one additional embodiment". The terms "first," "second," etc. may refer to different or the same object. Other explicit and implicit definitions may be included below.

Based on the relevant descriptions in the background art, it can be understood that the existing recognition model training methods can easily lead to the problem of recognition model aging during the recognition process, and often require repeated training multiple times to adapt to the continuous improvement of model recognition accuracy requirements and The continuous updating and iteration of identification object types will consume a lot of labor costs and time costs to continuously monitor and timely update the model aging. In order to overcome the above technical problems, in some embodiments provided by the present disclosure, Figure 1 shows a flow chart that provides an automatic iterative training method for a recognition model. The automatic iterative training method for a recognition model includes the following steps:

Step 101: Perform training operations on the recognition model based on the training data required for the current training round.

Step 102: Perform an inference operation on the recognition model that has completed the current training round to obtain the first recognition result containing pre-annotation information.

Step 103: Verify the first identification result to obtain the second identification result including verification information.

Step 104: Compare the first recognition result of the previous training round of the recognition model and the second recognition result of the current training round to generate training data required for the next training round.

Step 105: Repeat the above steps 101 to 104 to realize automatic iterative training of the recognition model. Among them, a preset model iteration time can be separated between two adjacent training rounds.

It can be understood that after each training round is completed, inference can be performed based on the current optimal model, and the pre-annotation result (that is, the above-mentioned first recognition result) is generated by the machine, and then the reviewer performs manual verification based on the annotation result. Review and fine-tuning (the result after review and fine-tuning is the above-mentioned second recognition result).

In the above embodiments, the technical solution provided by the present disclosure can be based on Pandas (a numerical computing extension tool based on the Python computer programming language, which incorporates a large number of libraries and some standard data models, providing efficient operation of large data sets The database implemented by the required tools) uses its relevant features to quickly and flexibly create the required training data, thereby greatly reducing maintenance and use costs; the database can also save and label results and manual review results at the same time and record the differences between the two. difference. The training performed by optimizing the training model each time a new training image is obtained is called a round of training. Then the optimized recognition model can generate a new training set based on the difference between the current recognition results and the previous round of annotation results. update iteration. In the specific implementation process, the model iteration time can be set so that the model itself can automatically iterate and update.

Furthermore, on the basis of realizing automatic iterative training of the recognition model, the training process of the model can be further optimized: It is understandable that in the existing technology, the quality of the recognition model is closely related to the quality of the manually labeled training data. However, manual annotation of training data has many problems such as high labor cost, heavy workload, and uneven annotation quality, which results in obvious differences in the effect of automatic iterative training of recognition models. In addition, existing image recognition models are often only used in general technical fields such as pedestrian detection and face recognition, and cannot meet the needs of ultra-large resolution defect detection and model adaptive iterative upgrades in specific technical fields.

In order to overcome the above technical problems, in some embodiments of the present application, Figure 2 shows a schematic flowchart of performing a training operation on a recognition model. As shown in Figure 2, the specific steps may include:

Train the recognition model through the optimized training set to generate a pre-trained model;

Optimize the pre-trained model so that the pre-trained model that meets the preset evaluation conditions is used as the recognition model to complete the current training round.

Step 201: Divide the image to be recognized to obtain the foreground area. It can be understood that by dividing the foreground area and the background area, the consumption of computing resources can be reduced, and at the same time, the interference of the complex background area on the foreground can be reduced. In the above step 201, a foreground segmentation model can be used to implement the segmentation operation. Those skilled in the art can select an appropriate segmentation method according to actual needs, which is not limited here.

Step 202: Perform first preset processing on the foreground area to obtain the minimum circumscribed rectangular marked area containing the target object. The specific implementation of the first preset processing will be explained in detail later.

Step 203: Perform second preset processing on the image to be recognized to obtain a preferred training set containing the minimum circumscribed rectangle marked area. The specific implementation of the second preset processing will be explained in detail later.

Step 204: Train the preset recognition model through the optimized training set to generate a pre-trained model.

Step 205: Optimize the pre-trained model so that the pre-trained model that meets the preset evaluation conditions is used as the recognition model that completes the current training round.

The automatic iterative training method of the recognition model provided in the aforementioned steps 201 to 205 can, on the one hand, overcome the problems of high labor cost, heavy workload, and inconsistent annotation quality of manual annotation. On the other hand, it can be applied to various specific application scenarios. Medium: In an application scenario of the present disclosure, the automatic iterative training method of the above recognition model can be applied to the inspection image recognition of wind turbine blades. It can be understood that in the process of using image recognition technology to inspect wind turbine blades, drones can be used to fly around the wind turbine blades and capture images of the surface of the wind turbine blades during the flight. Subsequently, the images can be captured by Image recognition is performed to obtain whether there are defects on the surface of wind turbine blades and the specific types of defects.

During the wind turbine blade inspection process, the image to be identified in the above embodiment may include an inspection photographed image containing the wind turbine blade, the foreground area may include the area occupied by the wind turbine blade in the inspection photographed image, and the target object may include the area present in the wind turbine blade. Defects on the blade surface. The specific implementation of the automatic iterative training method of the above recognition model will be explained and explained below using the application in wind turbine blade defect identification as an example.

In some embodiments of the present disclosure, further, before performing the first preset processing on the foreground area, the following steps may be included:

Determine whether the edge line of the foreground object in the foreground area can be identified and conforms to the preset characteristics: if so, continue to perform the first preset process; if not, directly obtain the marked frame containing the target object in the image to be recognized to generate training data.

It is understandable that during the process of wind turbine blade defect identification, if the most original blade foreground segmentation model makes a judgment error on a certain picture, the minimum circumscribed rectangular mark area cannot be obtained through the subsequent first preset processing. Therefore, it is necessary to determine in advance after completing the foreground division whether the edge lines of the foreground objects in the foreground area can be identified, and whether the edge lines of the foreground objects match the external features of the foreground objects. If any of the conditions are not met, the accurate minimum circumscribed rectangle marking area cannot be obtained. At this time, the marking frame containing the target object can be directly obtained from the original image to be recognized, and then the corresponding training data can be generated.

In some embodiments of the present disclosure, further, FIG. 3 shows a schematic flowchart of performing first preset processing on the foreground area. As shown in Figure 3, specific details may include:

Step 201: Obtain the edge line of the foreground object in the foreground area to obtain the angle of the foreground object in the image to be recognized. Taking defect identification of wind turbine blades as an example, when the foreground area is correctly divided, the location of the blades can be intuitively obtained through the edge lines of the wind turbine blades, and then the location of the wind turbine blades in the image to be identified can be calculated by extracting the edge lines. Approximate angle.

Step 302: Rotate the foreground area according to the angle so that the foreground object is in the vertical direction or the horizontal direction.

Step 303: Obtain the circumscribed rectangular mark box containing the target object in the rotated foreground area as the minimum circumscribed rectangular mark area.

It is understandable that most of the defects that appear on wind turbine blades are distributed along the direction of the wind turbine blades, and most of them are in a long, narrow, and tilted state; and when the wind turbine blades are in a tilted state in the image to be identified, they will face the long, narrow, and tilted state. When the defect is detected, if the marking frame wants to completely surround the defect area on the blade, a larger rectangular frame must be used. In the above step 202, the mask information corresponding to the foreground area can be used to rotate the foreground area so that the wind turbine blades are in the vertical direction or the horizontal direction. At this time, these narrow and inclined defects are also rotated to the vertical direction or the horizontal direction. The minimum enclosing rectangular marking area that can be obtained by marking the enclosing rectangular marking frame can further increase the proportion of defective foreground compared to directly marking the defect identification frame.

In some embodiments of the present disclosure, further, FIG. 4 shows a schematic flowchart of performing a second preset processing on an image to be recognized. As shown in Figure 4, specific details may include:

Step 401: Perform sliding window cropping on the image to be recognized according to the preset window size to obtain several cropped images that overlap with the minimum circumscribed rectangular mark area. It can be understood that in the usual artificial intelligence image recognition algorithm, the resolution of the input image can be 1333*800 pixels. In the application of wind turbine blade defect identification, the resolution of the captured pictures obtained during the inspection process exceeds 20 million pixels. Such huge captured pictures are obviously not suitable for direct input as images and need to be cropped through sliding windows. Perform preprocessing.

In the process of performing sliding window cropping, multiple fixed-size sliding windows can be set in the image to be recognized, and then the coincidence degree of each sliding window and the minimum circumscribed rectangle marked area is calculated one by one. If the coincidence degree is higher than the preset threshold, the The cropped image corresponding to the sliding window is used as qualified training data.

Step 402: Perform data enhancement processing on the cropped image to generate training data. Among them, the data enhancement process can use the mosaic algorithm to enhance the data, which is equivalent to increasing the number of batch images by 4 times, effectively saving computing resources. At the same time, based on the characteristics of the mosaic algorithm, random left and right inversions are performed during the data enhancement process. Random color conversion, random size scaling, random affine transformation, random rotation and other operations can effectively improve the generalization ability of the model. In the above embodiment, the four cropped images obtained after enhancement can finally be spliced to obtain the training data required for training.

In some embodiments of the present disclosure, further, in the process of training the pre-training model based on the training data, a method well known to those skilled in the art is adopted: that is, inputting the training data to the pre-built framework model, where The training data is the spliced image obtained after data enhancement in the above embodiment. In this embodiment, the pre-built framework model may be secondary developed based on the open source framework mmdetection (a deep learning target detection toolbox implemented by an open source Python machine learning library). The open source framework has basic versatility, but it is not perfectly suitable for all scenarios. Therefore, some local adjustments need to be made to the open source framework to meet the actual training needs and the rapid iteration and deployment requirements of possible subsequent needs. Those skilled in the art can complete the construction of the model as required based on their own knowledge and meet the training requirements of relevant application scenarios, which is not limited here.

In some embodiments of the present disclosure, further optimizing the pre-trained model includes the following steps: post-processing the recognition results of the pre-trained model using non-maximum merging.

It is understandable that in the automatic iterative training process of general recognition models, the post-processing method of Non-Maximum Suppression (NMS) is usually used. In the above-mentioned embodiments of the present disclosure, the post-processing method of non-maximum value merging is adopted in order to meet the actual application requirements of wind turbine blade defect identification. In the defect identification of wind turbine blades, it is required that the defect identification mark frame obtained after image recognition can all contain the area where the defect is located. However, the usual non-maximum suppression post-processing method may select a smaller one among multiple qualified candidate frames. The best option is used as the only choice to retain, thereby giving up other identification frames, which may result in the final selected identification frame still being unable to fully cover the single identification defect on the surface of the wind turbine blade; instead, multiple candidate frames are considered to jointly correspond to each other. Only under the condition of the largest external rectangle can the defect part to be identified be completely surrounded. Therefore, the use of non-maximum value merging to post-process the recognition results of the pre-trained model is a customized design for the above-mentioned specific application fields.

In some embodiments of the present disclosure, further, the preset evaluation condition includes that the recall rate of the preferred recognition model is greater than a preset threshold. It is understandable that in the process of evaluating image recognition models, the evaluation criteria often include two dimensions: recall and accuracy. Precision refers to the probability of identifying a target in an accurately recognized picture, while recall It refers to the ratio of the number of accurately identified targets to the number of targets in the training set.

In the application scenario of wind turbine blade defect identification, the recall rate is required to reach 99% or higher. In the above embodiment, the optimization process to achieve high recall rate includes:

Ensure that there is a sufficient number of positive samples for each defect type that is expected to be identified during model training; and

By adjusting the judgment threshold or judgment sensitivity for positive target detection results during the recognition judgment process.

Among them, in the process of ensuring a sufficient number of positive samples, it is necessary to perform more data enhancement work on some less common defect samples with higher severity levels, so that these defect types with a relatively small original sample size During the training process, it will not be affected by defect types with a particularly large sample size; in the process of adjusting the judgment threshold, you can check the probability distribution of the defect types in the predicted mark box by the model, and then appropriately reduce the prediction of true positives. The judgment threshold of the result. However, it is worth noting that lowering the prediction threshold will directly lead to more false positive predictions in the model, which will in turn affect the evaluation index of accuracy. In the application scenario of wind turbine blade defect identification, the ideal recall rate and accuracy evaluation criteria are that the recall rate is greater than 99% and the accuracy rate is greater than 45%.

In some embodiments of the present disclosure, FIG. 5 shows an automatic iterative training system for recognition models, which is applied to the automatic iterative training method for recognition models provided by the foregoing embodiments. Specifically, as shown in Figure 5, the automatic iterative training system of this recognition model can include:

The training unit 001 is used to perform training operations on the recognition model based on the training data required for the current training round.

The inference unit 002 is configured to perform an inference operation on the recognition model that has completed the current training round to obtain a first recognition result containing pre-annotation information.

The verification unit 003 is used to verify the first identification result to obtain the second identification result containing verification information.

The generation unit 004 is configured to compare the first recognition result of the previous training round of the recognition model and the second recognition result of the current training round to generate training data required for the next training round.

The iteration unit 005 is used to repeat the function implementation of the above-mentioned functional modules in order to realize automatic iterative training of the recognition model.

It can be understood that the functions implemented by the training unit 001 to the iteration unit 005 in the above functional module correspond to the operations performed in the aforementioned steps 101 to 105, and will not be described again here.

It can be understood that various aspects of the technical solutions of the present disclosure can be implemented as systems, methods or program products. Therefore, various aspects of the technical solution of the present disclosure can be specifically implemented in the following forms, that is, a complete hardware implementation method, a complete software implementation method (including firmware, microcode, etc.), or an implementation method that combines hardware and software aspects, which can be collectively referred to here as as "circuit", "module" or "platform".

FIG. 6 shows a schematic structural diagram of an electronic device according to some embodiments of the present disclosure. The electronic device is used to implement the automatic iteration method in the foregoing embodiments. The electronic device 600 implemented according to the implementation method in this embodiment will be described in detail below with reference to FIG. 6 . The electronic device 600 shown in FIG. 6 is only an example and should not impose any limitations on the functions and scope of use of any embodiment of the technical solution of the present disclosure.

As shown in Figure 6, electronic device 600 is embodied in the form of a general computing device. The components of the electronic device 600 may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different platform components (including the storage unit 620 and the processing unit 610), a display unit 640, and the like.

The storage unit stores program code, and the program code can be executed by the processing unit 610, so that the processing unit 610 executes the implementation of each functional module in the automatic iterative training system of the recognition model in this embodiment.

The storage unit 620 may include a readable medium in the form of a volatile storage unit, such as a random access unit (RAM) 6201 and/or a cache storage unit 6202, and may further include a read-only storage unit (ROM) 6203.

Storage unit 620 may also include a program/utility 6204 having a set of (at least one) program modules 6205 including, but not limited to: an operating system, one or more application programs, other program modules, and program data, Each of these examples, or some combination, may include the implementation of a network environment

Bus 630 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of a variety of bus structures. .

The audio and video signal synchronization processing device 600 can also communicate with one or more external devices 700 (such as keyboards, pointing devices, Bluetooth devices, etc.), and can also communicate with one or more devices that allow the user to interact with the electronic device 600, and /or communicate with any device (e.g., router, modem, etc.) that enables the electronic device to communicate with one or more other computing devices. This communication may occur through input/output (I/O) interface 650. Furthermore, the electronic device 600 may also communicate with one or more networks (eg, a local area network (LAN), a wide area network (WAN), and/or a public network, such as the Internet) through the network adapter 660. Network adapter 660 may communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in Figure 6, other hardware and/or software modules may be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tapes Drives and data backup storage platforms, etc.

In some embodiments of the present disclosure, a computer-readable storage medium is also provided. A computer program is stored on the computer-readable storage medium. When executed by a processor, the computer program can realize various functions in the automatic iteration system disclosed above. module implementation.

Although this embodiment does not exhaustively enumerate other specific implementations, in some possible implementations, various aspects described in the technical solution of the present disclosure can also be implemented in the form of a program product, which includes program code. When the program product When running on the terminal device, the program code is used to cause the terminal device to perform the steps described in the automatic iterative training method in the technical solution of the present disclosure according to the implementation methods in various embodiments of the technical solution of the present disclosure.

Figure 7 shows a schematic structural diagram of a computer-readable storage medium according to some embodiments of the present disclosure. As shown in Figure 7, a program product 800 for implementing the above method in an embodiment according to the technical solution of the present disclosure is described. It can adopt a portable compact disk read-only memory (CD-ROM) and include program code, and can be Run on terminal devices such as personal computers. Of course, the program product generated according to this embodiment is not limited to this. In the technical solution of the present disclosure, the readable storage medium can be any tangible medium that contains or stores a program. The program can be used by or in conjunction with an instruction execution system, device or device. In conjunction with.

The Program Product may take the form of one or more readable media in any combination. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: electrical connection with one or more conductors, portable disk, hard disk, random access memory (RAM), read only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.

A computer-readable storage medium may include a data signal propagated in baseband or as a carrier wave having readable program code thereon. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transport the program for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above.

The program code for performing the operations of the technical solution of the present disclosure can be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, etc., and also includes conventional procedural formulas. Programming language - such as C or similar programming language. The program code may execute entirely on the user's computing device, partially on the user's computing device, as a stand-alone software package, execute entirely on the user's computing device, partially on a remote computing device, or entirely on the remote computing device or server execute on. In situations involving remote computing devices, the remote computing device may be connected to the user computing device through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device, such as provided by an Internet service. (business comes via Internet connection).

In summary, through the technical solution proposed by the present disclosure, the training data required for the iteration of the recognition model can be automatically generated through the recognition results and verification results obtained between different training rounds, and these training data can be used to automatically upgrade the recognition model. Update iteration to avoid the situation where the recognition model ages and cannot meet the recognition needs. At the same time, it saves the labor cost of operation and maintenance upgrade of the recognition model, which has scalable value.

The above description is only a description of the preferred embodiments of the disclosed technical solution, and does not limit the scope of the disclosed technical solution. Any changes or modifications made by those of ordinary skill in the field of the disclosed technical solution based on the above disclosure shall belong to the claims. scope of protection.

Claims (10)

An automatic iterative training method for a recognition model, which is characterized by including the following steps: Perform training operations on the recognition model according to the training data required for the current training round; Perform an inference operation on the recognition model that has completed the current training round to obtain a first recognition result containing pre-annotation information; Verify the first identification result to obtain a second identification result containing verification information; Comparing the first recognition result of the previous training round of the recognition model and the second recognition result of the current training round to generate training data required for the next training round; Repeat the above steps to achieve automatic iterative training of the recognition model. The automatic iterative training method of recognition model according to claim 1, characterized in that a preset model iteration time is separated between two adjacent training rounds. The automatic iterative training method of a recognition model according to claim 1, wherein the recognition model is used to recognize an image to be recognized containing a target object to obtain the identification information of the target object; The training data includes the image to be recognized. The automatic iterative training method of the recognition model as claimed in claim 3, wherein the training operation on the recognition model includes the following steps: Divide the image to be recognized to obtain the corresponding foreground area; Perform first preset processing on the foreground area to obtain the minimum circumscribed rectangular mark area containing the target object; Perform a second preset process on the image to be recognized to obtain a preferred training set containing the minimum circumscribed rectangle marked area; Train the recognition model through the preferred training set to generate a pre-trained model; The pre-trained model is optimized to use the pre-trained model that meets the preset evaluation conditions as the recognition model that completes the current training round. The automatic iterative training method of the recognition model according to claim 1, wherein the first preset processing of the foreground area includes the following steps: Obtain the edge line of the foreground object in the foreground area to obtain the angle of the foreground object in the image to be recognized; Rotate the foreground area according to the angle so that the foreground object is in a vertical direction or a horizontal direction relative to the image to be identified; A circumscribed rectangular mark box containing the target object is obtained in the rotated foreground area as the minimum circumscribed rectangular mark area. The automatic iterative training method of the recognition model according to claim 1, wherein the second preset processing of the image to be recognized includes the following steps: Perform sliding window cropping on the image to be identified according to a preset window size to obtain a number of cropped images that overlap with the minimum circumscribed rectangular mark area; Perform data enhancement processing on the cropped images to generate the preferred training set. The automatic iterative training method of the recognition model as claimed in claim 1, wherein the optimization of the pre-trained model includes the following steps: The recognition results of the pre-trained model are post-processed using non-maximum value merging. An automatic iterative training system for recognition models, characterized in that it is applied to the automatic iterative training method of recognition models according to any one of claims 1 to 7; The automatic iterative training system of the recognition model includes: A training unit, configured to perform training operations on the recognition model based on the training data required for the current training round; An inference unit, configured to perform an inference operation on the recognition model that has completed the current training round to obtain a first recognition result containing pre-annotation information; A verification unit, configured to verify the first identification result to obtain a second identification result containing verification information; A generation unit configured to compare the first recognition result of the previous training round of the recognition model with the second recognition result of the current training round to generate training data required for the next training round; An iterative unit is used to repeat the above steps to realize automatic iterative training of the recognition model. An automatic iterative training device for recognition models, which is characterized by including: Memory, used to store computer programs; A processor, configured to implement the automatic iterative training method of the recognition model according to any one of claims 1 to 7 when executing the computer program. A computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the recognition model as described in any one of claims 1 to 7 is implemented. An automatic iterative training method.

Images